When GPU Resources Run Dry, It Still Looks Like Everything Is Working: A Tale of Two AI Competing for VRAM

I was running video lip-sync generation and pose estimation from video as separate AI sessions on the same machine. Both slowed down, and one of them started outputting a 'success-looking failure'—taking 431 seconds per frame while detecting zero data. With VRAM usage and system load maxed out, it appeared to be running perfectly fine at first glance.